CW2 Training Instructions

CW2_training_instructions

User Manual:

Open the PDF directly: View PDF .
Page Count: 3

Instructions for training/testing of your designs:

Dataset:

CUHK03 is a very popular dataset containing 13,164 images of 1,360 pedestrians. Sample images

from this dataset are presented below:

For the purpose of this coursework you are provided with both images from the dataset and feature

vectors extracted from the images. Images in this dataset were captured with use of two cameras,

what is more, each person was photographed a number of times. Therefore, each identity (person) is

represented in the dataset approximately 8-9 times. All the data you need is contained in

CW2_data.zip. Please unpack it, and you should find the following contents:

1. Folder images_cuhk03 contains all the images from the dataset. Their names may seem illogical

at first, but they are properly annotated in the file cuhk03_new_protocol_config_labeled.mat.

2. Matlab file cuhk03_new_protocol_config_labeled.mat. Once you open this file with Matlab or

Python, you will find 6 main components (in the form of arrays):

a. camId specifies whether image was taken from camera 1 or camera 2. While evaluating

(testing) your algorithms, you should not consider images of your current query identity taken from

the same camera. For example, when you create ranklist for the first query image (index 22, label 3,

camera 1, name "1_003_1_02.png"), you should not include images with indexes 21, 23, 24 in this

ranking list, as these are images of the same person (label 3) captured by the camera with index 1.

b. filelist specifies correspondences between names of files in images_cuhk03 and their

indexes.

c. gallery_idx, which specifies indexes of the part of the dataset from which you compose

your ranklists during testing phase

d. labels contains ground truths for each image

e. query_idx contains indexes of query images

f. train_idx contains indexes of images that can be used for training and validation

To load cuhk03_new_protocol_config_labeled.mat into Python you can use the code from this

example (and repeat for other arrays listed above):

from scipy.io import loadmat

train_idxs = loadmat('cuhk03_new_protocol_config_labeled.mat')

['train_idx'].flatten()

with Matlab you can simply use load('cuhk03_new_protocol_config_labeled.mat') or double click

on the file in File Explorer.

Apart from two files mentioned above, file feature_data.json contains an array 14096x2048 with

some feature vectors representing each image in the dataset. YOU SHOULD PERFORM YOUR

COMPUTATIONS WITH USE OF FEATURE VECTORS PROVIDED, IMAGES ARE

THERE ONLY FOR THE REFERENCE AND VISUALISATION. With Pyhon you can simply

load it as follows:

import json

with open('feature_data.json', 'r') as f:

features = json.load(f)

and with Matlab:

features = jsondecode(fileread('feature_data.json'));

Please note it takes a while to load a file as it contains considerably large amount of data.

Training, validation, testing:

1. Training:

For the purpose of training use feature vectors provided in the file feature_data.json. Indexes of

vectors that can be used for training are stored in the train_idx vector belonging to

cuhk03_new_protocol_config_labeled.mat.

2. Validation:

You can use 100 randomly selected identities from training set as validation set. Please note each

identity is represented in the dataset 7-10 times (there is a number of photos/feature vectors of each

person in the dataset), and you should include all of them in the validation set, therefore around

700-1000 different feature vectors will be used for validation. There exists an idea in Computer

Vision that you use validation set only to specify the number of iterations that is optimal for your

design and then you include your validation set into train set and perform final training (without

validation set) for this fixed amount of iterations. You can try this idea as well. Once again, use

feature vectors provided in the file feature_data.json. Indexes of vectors that can be used for

training/validation are stored in the train_idx.

3. Testing:

Finally, you perform testing. Both query_idx and gallery_idx specify your test set. The task is to

take all the feature vectors from query set, one by one, and create a ranklist for each of them, by

finding nearest neighbours within gallery_idx, deleting images/feature vectors of considered query

identity, captured by the same camera (same label and same camId). Sample ranklists after

deletion, with use of dataset images for visualisation (green frame indicates positive matches, red

are negative ones, black are query images, you can see positive matches presenting different views

of the query identity, due to deletion mentioned above):

Once having your ranklists you can calculate @rank1, etc. scores for such ranklists, achievable

scores for simple nearest neighbour and Euclidean distance are:

top1:0.469286 top5:0.668571 top10:0.750000 mAP:0.432923 (I managed to do a bit better with

simple NN on Matlab - @rank1 = 0.47, so don’t worry if you achieve a bit better scores).

Please do not mix sub-datasets:

train_idx: training, validation (optional)

query_idx + gallery_idx: only for testing your design, follow the procedure above

References:

W. Li, R. Zhao, T. Xiao, and X. Wang, “Deepreid: Deep filter pairing neural network for person re-

identification,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition, June 2014,

pp. 152–159.

Z. Zhong, L. Zheng, D. Cao, and S. Li, “Re-ranking person re-identification with k-reciprocal

encoding,” CoRR, vol. abs/1701.08398, 2017. [Online]. Available: http://arxiv.org/abs/1701.08398

CUHK03 dataset:

http://www.ee.cuhk.edu.hk/~xgwang/CUHK_identification.html

https://github.com/zhunzhong07/person-re-ranking/tree/master/CUHK03-NP

CW2 Training Instructions